Tracing the ghost of the 2017 contract never gets easier. Every cycle, the shape returns: a brilliant person, a compelling story, a wager that grows until it consumes the wagerer. In July, a hedge fund called Situational Awareness — run by Leopold Aschenbrenner, a former OpenAI researcher who once argued that artificial general intelligence would arrive by 2027 — lost roughly $35 billion in a leverage-driven collapse. Assets fell from a peak above $45 billion to around $10 billion. An approximate 78% drawdown. Now, according to reporting that first moved through CNBC and then echoed across crypto-Twitter, the same fund is back — not with borrowed spot, but with call options on five names: AMD, SK Hynix, SanDisk, CoreWeave, and Bloom Energy.
That is the hook. Not "a man bets on AI." Something stranger: a person who built his public identity on a specific timeline for machine superintelligence has now converted his conviction into contracts with expiration dates. The canvas shifted, but the buyer remained — only now he is paying rent on his own faith. For those of us who spent the last decade watching crypto narratives metabolize hope into leverage and back again, this is not a foreign story. It is our story, wearing a different jacket. And it matters because the five tickers are not random. They are a map. Read together, they describe where the next bottleneck of the entire intelligence economy will sit — and, by extension, where the crypto assets that mirror that economy are heading.
The reporting itself is thin. We are working from an anonymous-source CNBC story, amplified by a tweet from the investor Shay Boloor, and re-narrated by a crypto outlet with an obvious incentive to ride the AI news cycle. There is no 13F. There is no official disclosure. There is a subpoena from the SEC aimed at the fund's dealings with Wall Street banks, and there is a discount purchase of the fund's distressed positions by Ken Griffin's Citadel. That is the raw material. Everything below is forensic reconstruction — and I will flag clearly where I am inferring rather than reporting.
Context
Let me place the man properly, because the details matter more than the drama.
Leopold Aschenbrenner is not a generic fund manager. He is a specific archetype: the safety researcher who became an accelerationist. He worked on OpenAI's Superalignment team — the group tasked with ensuring that a superintelligent system would not destroy us — before leaving in 2024 and publishing Situational Awareness: The Decade Ahead, a widely read series arguing that the race to AGI was entering its final, decisive phase, and that the nations and labs which recognized this would be transformed. The essay was, in effect, a narrative artifact. It described a world where compute, energy, and intelligence converge, and it gave that world a deadline.
That deadline — 2027 — is the key. Every investment thesis has a temporal spine. Value investors say eventually. Momentum traders say now. Aschenbrenner's public framework said by 2027, which is an unusually precise claim for a field as noisy as machine learning. When you attach a date to a belief, you convert faith into a contract. And contracts, unlike faith, can expire worthless. This is the single most important thing to understand about the July collapse and the new position: a man whose entire intellectual brand rests on a clock has finally started trading like the clock is real.
The fund's architecture followed the thesis. Situational Awareness raised capital on the premise that the years before AGI would be the most lucrative investment window in modern history — that the infrastructure of intelligence, not the intelligence itself, would capture the value. That is a defensible argument. It is also the argument every capital allocator in 2024 and 2025 was making, which is precisely why the position became crowded.
Mapping the invisible liquidity flows of summer helps here. In June and July, the AI trade was the consensus trade. Every fund — from multi-manager pods to family offices — held some version of it. When a trade becomes consensus, it stops being a bet and becomes a position: you are no longer wagering, you are holding, and holding through a drawdown requires the one thing leverage destroys — time. Aschenbrenner's leverage removed his time. The margin call is a clock that rings faster than any expiry.
So the July collapse was not an anomaly. It was the natural end-state of a faith expressed through borrowed money. The question the new options position raises is subtler: has he learned, or has he simply re-engineered the same error with a floor beneath it? Because a call option is not a confession of humility. It is a hedge on being wrong about timing — and nothing else.
I have watched this exact conversion before. During the 2017 ICO sprint, I audited fifteen whitepapers for a small Austin-based venture group, and I noticed that the projects which survived the subsequent winter were not the ones with the best technology. They were the ones whose founders had quietly switched from narrative to structure — from promising a future to financing it one bounded bet at a time. The ones who kept promising infinity went to zero. The ones who started pricing their own conviction — that is, who began buying time with defined risk — persisted. Aschenbrenner has just done a version of this. Whether he did it consciously or was forced into it by a Citadel fire sale is the open question. The market does not much care about intent. It cares about structure. And the structure changed.
Core: The Five-Contract Map
Here is where the crypto reader should lean in, because Aschenbrenner's five tickers are not a stock basket. They are a thesis expressed as a supply chain. And a supply chain is the one thing the crypto industry has spent five years trying to disintermediate, tokenize, and occasionally destroy. Read the list again slowly:
- AMD → compute silicon (the only scaled challenger to NVIDIA's AI accelerator dominance)
- SK Hynix → HBM, high-bandwidth memory (the binding constraint on every AI chip)
- SanDisk → NAND/SSD storage (the substrate for training data and inference caches)
- CoreWeave → GPU cloud (the wholesale layer of AI compute)
- Bloom Energy → fuel cells and data-center power (the emerging physical limit)
Computation. Memory. Storage. Cloud. Energy. That is not a portfolio. That is a stack. Five layers, one demand driver: AI capital expenditure. Which means it is not diversification at all — it is the same risk factor wearing five costumes. This is the single most important structural observation in the whole story, and almost every retelling of the CNBC piece missed it, because most financial journalists analyze tickers, not systems.
I learned to read bets this way during DeFi Summer, when I mapped $2.3 billion in total value locked across Aave and Compound and realized that "the money lego narrative" was really a single leverage cycle dressed in a dozen protocol names. When the cycle reversed, every "diversified" DeFi portfolio fell together, because the assets shared one collateral. Aschenbrenner's five contracts share one collateral too: the assumption that compute demand grows without limit. If that assumption holds, all five float. If it breaks, all five sink — and options, unlike spot, sink on a schedule.
Let me take the layers one at a time, because the interesting information is in the edges — in where crypto and AI infrastructure are converging, and where crypto has already missed the boat.
The Compute Layer: AMD and the Challenger's Premium
Aschenbrenner did not buy NVIDIA. He bought AMD. That omission is louder than any of the inclusions.
There are three plausible readings, and I would bet on a blend of all three. First, NVIDIA's valuation already reflects near-total victory — the option premium on further upside is thin, and a fund rebuilding after a 78% drawdown needs convexity, not safety. Second, AMD carries a "catch-up premium": if MI-series accelerators close even a fraction of the software-ecosystem gap with CUDA, the re-rating is violent. Third — and this is where crypto readers should pay attention — betting on the challenger is a narrative bet, not a fundamentals bet. It is the same instinct that made people buy the second Layer 2 instead of the first, the alternative L1 instead of Ethereum, the fork instead of the original. The challenger's story is always more elastic, because it has somewhere to go.
But here is the trap, and it is the trap the decentralized-compute sector keeps falling into. A challenger's premium is real only if the moat is software, not silicon. CUDA is not a chip; it is two decades of accumulated developer habit. Habit, in computing, is stickier than performance. This is precisely the lesson the decentralized GPU networks — Render, Akash, io.net and their cousins — have been forced to learn the hard way. They can rent you cheaper compute. They cannot rent you the twelve years of tooling, libraries, and graduate-student muscle memory that make that compute usable for frontier training. The decentralized networks have largely converged on inference and rendering not because they lack ambition, but because the moat lives upstream of the hardware.
So AMD is the equity market's version of the same wager the DePIN sector has been making: that the incumbent's dominance is a habit, and habits can be broken by a sufficiently motivated challenger. Sometimes they can. Usually they cannot. The interesting question is not whether AMD's flagship part benchmarks well. It is whether the developer abandons CUDA. And developers abandon tools for one reason only: because their employer forces them to, to save money at scale. Watch the hyperscalers. They are the ones who will decide whether the challenger camp is real.
The Memory Layer: SK Hynix and the Constraint Crypto Forgot
If AMD is the narrative bet, SK Hynix is the physical bet, and it is arguably the sharpest position in the entire basket.
Here is the mechanism. AI accelerators are not limited by their arithmetic. They are limited by how fast they can be fed. HBM — high-bandwidth memory — is the straw through which data reaches the compute cores, and for the last two years the straw has been the bottleneck. HBM is sold out. HBM capacity is being rationed. HBM4 is the next generational step, and only a handful of firms on Earth can make it at scale. When you cannot get enough HBM, it does not matter how many GPUs you ordered. This is a physical chokepoint, not a financial one, and physical chokepoints are the most durable margins in any supply chain.
This is exactly the dynamic the crypto industry has spent years trying to create — and mostly failed to. When I audited NFT collections in 2021, I categorized a thousand projects by cultural capital and found that the ones with genuine scarcity — a capped supply tied to a real social mechanism — outlasted the ones with infinite mints. HBM is the industrial version of that insight. Scarcity that cannot be inflated. Supply that cannot be voted into existence by a DAO. In a world where everything digital can be copy-pasted, a physical process running at the edge of human manufacturing capability is the closest thing to honest scarcity we have. That is why the HBM vendors can price like monopolists and why the position makes sense.
But — and this is the part the celebratory retellings skip — every durable bottleneck eventually invites capacity. SK Hynix and its competitors are pouring capital into HBM expansion. The same reflex that turns a shortage into a windfall turns a windfall into a glut. The critical variable to track is not whether HBM is scarce today; it is whether the expansion outruns the demand by some quarter in 2026 or 2027. When I mapped the DeFi lending markets, I watched the same pattern: yield that looked structural was really a temporary imbalance, and when the imbalance closed, the yield vanished overnight. Aschenbrenner is betting on the persistence of the memory bottleneck. Historically, that has been a losing bet. Bottlenecks get solved. It just takes longer than the bears think.
The Storage Layer: SanDisk and the Invisible Substrate
SanDisk — the NAND and SSD business — is the quietest and most underrated leg. Nobody tweets about storage. Nobody writes threads about flash memory. Which is precisely why it is interesting.
AI is not primarily a computation problem at the infrastructure level. It is a data movement problem. Models must be trained on petabytes, checkpointed constantly, and served to users with low latency. Every one of those operations hits storage. The inference layer — the part everyone now assumes will dominate as models are deployed rather than trained — is essentially a cache-modelling problem: how fast can you fetch the relevant piece of data to feed the next token?
I have always found that the most valuable positions are the ones that sit beneath the narrative, in the substrate nobody discusses. When I worked through the 2022 collapse of FTX's narrative trust, I audited more than fifty venture funding announcements and found that the projects which preserved value were rarely the most talked-about ones. They were the ones whose function was infrastructural — the ones whose absence would be felt as a friction rather than a headline. Storage is that. You do not notice it until it is gone.
For crypto readers, there is a specific echo here: the decentralized storage networks — Filecoin, Arweave and their successors — have spent years arguing that AI training data needs a censorship-resistant, verifiable home. This is a real thesis with a real, if slow, traction curve. But note what Aschenbrenner did not do: he did not buy a decentralized storage token. He bought a legacy NAND manufacturer. That is the market's verdict on where durable storage economics currently live — and it is a verdict the crypto storage sector should take seriously rather than deflect. Decentralized storage has a values proposition. It has not yet built an economics proposition that competes with a fab. Until it does, the substrate belongs to the incumbents.
The Cloud Layer: CoreWeave and the Decentralized Shadow
CoreWeave is the most aggressive leg of the basket. It is a "pure AI cloud" — a company whose business is essentially renting GPU capacity at wholesale, tightly coupled to NVIDIA's supply. Buying CoreWeave calls is buying the purest expression of compute rental demand growth. No legacy business drags down the story. No diversification dilutes the bet. It is maximum beta, maximum elasticity, maximum downside.
This is where the crypto parallel becomes almost too on-the-nose. The entire DePIN narrative — decentralized physical infrastructure networks — is a bet on the same thing: that compute rental demand will grow faster than centralized provisioning can satisfy it, and that distributed supply can capture the overflow. CoreWeave is the centralized version of that thesis. Render, Akash, io.net are the decentralized version. They are competitors in the same logical space, and the market is currently pricing the centralized version at a premium while the decentralized version trades as a long-tail speculation.
That is a curious asymmetry, and it deserves a moment of honest analysis rather than reflexive crypto cheerleading. Why does the centralized GPU cloud command a premium? Because it offers reliability, SLAs, and a single throat to choke. An enterprise training a frontier model does not want to orchestrate an amorphous swarm of consumer GPUs. It wants a contract, a phone number, and a guarantee. The decentralized networks solve a cost problem. They do not yet solve a trust problem. And in infrastructure, trust is the product.
Now the risk. CoreWeave's business is, structurally, a leveraged bet on the persistence of compute scarcity and on NVIDIA's continued willingness to allocate supply. It has enormous elasticity precisely because it has enormous concentration. If compute supply catches up to demand — if the frontier labs slow their scaling, if efficiency gains compress the required GPUs per unit of intelligence — the pure-play cloud is the first to feel it. This is why buying CoreWeave calls after a leverage blowup is, to put it gently, an aggressive act. It is not a hedge. It is the same trade, re-instrumented.
The Energy Layer: Bloom Energy and the Miner's Mirror
And now the most interesting leg of all — the one that tells you where the thesis is heading.
Aschenbrenner bought Bloom Energy: solid-oxide fuel cells, essentially on-site power generation. He did not buy a utility. He did not buy a power grid operator. He bought the distributed answer to the question of how you power an AI data center when the grid takes three to five years to expand and your compute deployment takes months.
This single position tells you more about Aschenbrenner's real thesis than the other four combined. It says: the binding constraint on AI is no longer chips. It is electrons. The compute layer can be purchased. The memory layer can be expanded. But the power layer is governed by physics, permitting, and the speed of human construction — variables no amount of capital can instantly accelerate. The next bottleneck of the intelligence economy is not silicon. It is electricity. That is the real insight buried in this basket, and it is the one I would carry forward even if I disagreed with every other position.
Here the crypto industry has a story that is genuinely instructive, and I want to tell it carefully because it is often told badly.
Over the last two years, Bitcoin miners have quietly begun converting their facilities into AI and high-performance-computing hosting. The logic is almost too clean: a miner already owns land, power contracts, and cooling infrastructure — the three things an AI data center needs and the three things that take the longest to procure. What a miner lacks is the compute itself. So the deal writes itself: bring the GPUs, use my power.
Summer taught us that liquidity has a heartbeat, but it turns out that power has a heartbeat too, and it beats on a much slower rhythm — the rhythm of interconnection queues and turbine deliveries. The miners understood this before almost anyone else, because their entire industry lives or dies on the cost of a kilowatt-hour. When the AI capex wave arrived and the grid proved unable to absorb it, the miners discovered they were sitting on the scarce asset. Some of them have re-rated dramatically on this pivot alone.
So when Aschenbrenner buys a fuel-cell company, he is not buying an exotic niche. He is buying into the same physical insight that is quietly rewiring the Bitcoin mining sector into an AI power-and-shell business. The two industries — crypto mining and AI infrastructure — are collapsing into each other, and the collapse is being driven by one shared constraint: the speed at which you can put a megawatt behind a workload. This is the single most under-discussed convergence in the entire market, and it is where I would tell an institutional client to look if they wanted exposure that is both crypto-native and AI-levered, without taking the crowded equity trade.
The Sentiment Machine: Two Narratives, One Clock
Every codebase is a whispered promise, but so is every ticker — and in 2025, the market was listening almost exclusively to the AI promise. That is the meta-story underneath this whole event, and it matters for anyone trading crypto narratives.
Here is what I tracked. When I launched my exploration into AI-agent trading in 2026, and ran two narrative-detection bots across 10,000 AI-generated tweets, the finding that rattled me was not the volume. It was the speed. AI-driven narratives compressed market cycles by roughly forty percent. A story that used to take a month to build, peak, and decay was now completing in eighteen days. The mechanism is obvious in retrospect: the same machine intelligence that lets a fund size a position in seconds also lets a narrative propagate in hours. The informational velocity of the system had risen, and velocity, in any system, is a proxy for fragility.
This is why Aschenbrenner's August collapse was so violent. He was not merely leveraged. He was leveraged inside a narrative that had itself become fast. When the AI story wobbled in July, the propagation was near-instantaneous. There was no slow bleed to react to. There was a cliff.
Now map that onto crypto. The AI narrative and the crypto narrative are, at this moment, in a peculiar relationship. They share a supply chain (energy, fabs, capital), they share an investor base (the same risk-hungry allocators), and — increasingly — they share a clock. When one narrative runs hot, it borrows liquidity from the other. The AI trade has spent eighteen months systematically out-ranking crypto for the attention of the generalist allocator. That is not a moral judgment about the technologies. It is a flow fact. And flows, in the short run, are the only thing that moves price.
The five-contract basket is, in this reading, a sentiment instrument as much as a position. It is an expression of where a specific, sophisticated, borderline-ideological subset of capital believes the clock is pointing. I would not follow it. But I would absolutely read it — the way a geologist reads a fault line. It tells you where the stress is accumulating.
The Regulatory Overlay
There is one more layer, and it is the one that determines whether this story has a second act.
The SEC has subpoenaed the fund over its dealings with Wall Street banks. That word — subpoena — is doing a lot of work in a sentence that also contains the phrase re-bet on AI stocks. We are talking about a fund whose distressed positions were bought at a discount by Citadel, whose leverage blew up, and whose survival is now partly a matter of how much regulatory friction it can absorb. Against that backdrop, the cheerful "he is back!" framing of the CNBC story is almost comically detached from the risk.
And here I will state something I have believed for years, because this is the case that proves it. Most KYC is theater. I have said this before and I will say it again: the compliance apparatus imposed on honest users is, in practice, a toll booth that determined actors route around at trivial cost — a few wallets here, a shell entity there — while the compliant majority pays the full price in friction, delay, and foregone access. Now watch the same pattern at institutional scale. The banks that facilitated this fund's leverage are the very banks now under subpoena. The compliance layer did not prevent the blowup. It merely determined who was left holding the paperwork when it happened. Compliance costs are always passed, in the end, to the honest — never to the clever. This is not cynicism. It is a pattern I have observed across every regulatory regime I have mapped since 2018, and it is reconfirmed every single cycle.
If the SEC inquiry escalates, the fund's ability to raise fresh capital — never high after a 78% drawdown — collapses further. Which brings us to the inference I have been holding back.
Contrarian: The Less Dramatic Reading
Everyone is reading this as a redemption arc — the fallen believer rising again with a cleverer instrument. I am not convinced. Here is the reading the coverage avoided.
The CNBC story never says where the money came from. It explicitly notes that it is unclear whether the options were funded with remaining cash or with new investor capital. I have a strong prior on which it is, and the prior is not flattering. A fund that has just shed 78% of its value, is under SEC subpoena, has had its distressed positions bought at a discount by a vulture, and is run by a manager whose stated thesis carries a 2027 deadline — that fund does not raise a fresh round easily. The far more plausible inference is that this is residual capital being redeployed by a manager who cannot admit the thesis is on a clock. That is not a renewal. That is a burn.
Now the deeper contrarian point. Everyone treats the shift from leverage to options as evidence of learning — an embrace of defined risk. I think the shift is evidence of the opposite: an embrace of a deadline. Leveraged spot is the instrument of someone who believes he can hold forever. Options are the instrument of someone who has been forced to admit that forever is not available. The structural change is not "he got cautious." It is "he got time-bounded." He has finally started trading his own thesis like a thesis — with an expiry. The strike price he chose is really a confession of the deadline he spent years refusing to name.
And there is a final contrarian cut. The reason this story has any news value at all is not that the bet is clever. It is that the bet is dramatic. A man who lost $35 billion and then re-bet is a character. A fund that did the quiet, boring thing — wound down, returned capital, waited — would never have been written about. The market rewards the narrative of the re-bet, which is the exact dynamic Aschenbrenner himself predicted: that in the decade ahead, stories, not fundamentals, would move capitals. He has become a case study in his own thesis, and I am not sure that is a compliment.
There is also a detail almost no one mentioned. He kept his private positions — Anthropic, among others. That matters enormously. It tells you he has not abandoned his long-term AGI conviction. He has simply stratified it: the private book holds the faith, the public book holds the trade. That is actually the most rational thing he has done in this entire saga. Faith belongs in instruments that can wait. Trades belong in instruments that cannot. If he had done that from the beginning — kept the religion in private equity and the timing bet in defined-risk public markets — there would be no July to speak of. The disaster was never the belief. It was the instrument the belief was poured into.
Takeaway
So where does this leave us, standing at the edge of the next cycle?
Watch the energy leg, not the chip leg. The position that will be remembered from this basket is Bloom Energy, because it points at the constraint that will outlast every chip cycle: electricity, and the impossibility of conjuring a megawatt on demand. The AI-crypto convergence I have been tracking is, at bottom, a power convergence. The miners figured this out first. The market is only now catching up.
Watch the 13F. If a formal filing appears, the inference chain in this piece either holds or breaks — and it will break publicly, in a document, which is more than the CNBC sourcing gave us. Watch, too, whether the challenger camp delivers on software. If AMD's ecosystem closes even a fraction of the gap with CUDA, the entire challenger thesis re-rates — and the decentralized-compute networks will feel it in both directions.
But the real thing to watch is the clock. Two years ago, the market priced AGI as an eventuality — no date, no discount, no expiry. Now, a manager whose entire brand was built on a specific year has started buying time-bounded contracts. When the believers start hedging their own timelines, the narrative has already begun to shift beneath them. We were swimming in a sea of narrative, and Aschenbrenner has just thrown us the first life raft — himself. The question every allocator should be asking is not whether he is right about AI. It is whether the clock he finally acknowledged is his alone, or whether it is the market's. Because if it is the market's, then the option premium on AGI is already being priced — and most of us are still buying the stock.
